Probabilistic forecasting of regional solar power incorporating weather pattern diversity

Power grid stability depends on the ability to forecast solar power generation at the regional level. Most previous research on probabilistic forecasting has focused on the use of machine learning to predict the output of individual solar power plants rather than regional solar power generation, and few studies have considered the effects of seasonal weather patterns at the regional level. In this study, climate and geographic data were collected from 83 weather stations between 2019 and 2021 for use in developing a probabilistic model by which to predict regional solar power generation. The results of weather pattern analysis based on unsupervised machine learning and ensemble voting were used to build a probabilistic quantile regression model for the shortterm prediction of regional solar power generation capacity. The efficacy of the model was assessed using data from 48 solar PV power plants, which included four sub-datasets pertaining to four target regions. Highly accurate results were consistently obtained across all regions in both the winter and summer seasons. The proposed probabilistic model outperformed conventional deterministic model by 6.55% and conventional probabilistic model by 4.03% in terms of total normalized mean absolute error (NMAE). Prediction intervals generated by the proposed model could be used as input parameters for regional power dispatch decisions.